ORACIS
Start a project

AI app development · LLM applications · AI agents

Sheet
A-050
Set
Practice
Status
Issued

AI app development, from LLM features to AI agents.

We build AI into products: LLM features inside the software you already run, AI agents with real limits, and new AI-native products. Each one is measured against the job it has to do, capped on cost, and reviewed by a person where the stakes are high.

Discuss your AI project How we build it →

A-050PracticeIssued

How an AI feature reaches production A user request goes to a model call, several models at a capped cost, then to an evaluation that scores it before it ships, then to a human review where the stakes need it, and only then to a response. A call that fails the check takes a dashed path to a safe fallback. User request Model call several models, capped cost Evaluation scored before it ships Human review where the stakes need it Response fails the check Safe fallback How an AI feature reaches production A user request goes to a model call, several models at a capped cost, then to an evaluation that scores it before it ships, then to a human review where the stakes need it, and only then to a response. A call that fails the check takes a dashed path to a safe fallback. User request Model call several models, capped cost Evaluation scored before it ships Human review where the stakes need it Response fails the check Safe fallback
Fig. D-1An AI feature is evaluated before it answers, and a person decides where it matters

What we build

AI development services, from one feature to a whole product.

01

LLM features in your app

Assistants, drafting, extraction, search and summaries built into software you already run, with real limits on latency, cost and failure, not bolted on as a demo.

02

AI agents

Agents that use tools and take steps on their own, inside limits: which tools they may call, what they may spend, and which actions wait for a person to approve.

03

AI-native products

New products where the model is the core: architecture, data pipelines, prompt and tool design, and the production systems around them.

04

Answers from your own documents

Search and question-answering over your files, tickets or records, with each answer tied back to the source it came from so a reader can check it.

05

Evaluation

Test sets and graders that measure the feature against the job it has to do, so changing a model or a prompt is a measured decision, not a hunch.

06

AI in regulated workflows

Clinical, legal and financial work, where the founding team has shipped AI that a licensed professional reviews before anything is relied on.

How it works

Measured before it ships, with a person where it matters.

A demo that answers well five times is easy. An AI feature that answers well on the five-hundredth unusual request, at a cost you can predict, is the actual work. That is why the test set comes before the build.

  1. Discovery

    A short, priced discovery: what the AI has to do, what a wrong answer costs, what data it may see, and whether a model is the right tool at all.

    You get A build plan with the approach, the risks and a fixed quote.

  2. Evaluation set before the build

    We write down what good looks like as test cases and graders, before the feature is built, so progress is measured from the first day.

  3. Build

    The feature, its guardrails and its fallbacks, built in your codebase with every change reviewed. Deterministic rules wherever a model's judgment is not enough.

  4. Launch and watch

    Cost limits, logging and monitoring in production, with the evaluation re-run whenever a model or prompt changes.

    You get An AI feature you can measure, afford and switch off safely.

Shipped by the founding team

Multi-model review, with a deterministic safety backstop.

Clinical AI output is checked by several independent models in different roles. Disagreement goes to a clinician, and a rules engine, not a model, decides what reaches a patient. In production today.

Multi‑model review with a deterministic backstop A lab result goes to three independent models in different roles. If they disagree it escalates to a clinician. If they agree, a deterministic rules engine decides whether anything is released to the patient, and can block it. Lab result Interpret Challenge Cross-check agree? disagree → clinician Rules engine deterministic release block Multi‑model review with a deterministic backstop A lab result goes to three independent models in different roles. If they disagree it escalates to a clinician. If they agree, a deterministic rules engine decides whether anything is released to the patient, and can block it. Lab result Interpret Challenge Cross-check agree? disagree agree Clinician human review Rules engine deterministic block release
Fig. W-03No single model output reaches a patient unchecked
Section-by-section attestation AI drafts four sections of a clinical note. A provider attests each one. Three are attested and one is not, so signing stays locked at three of four. AI draft History Assessment Plan Orders Sign note locked · 3 / 4 Provider attests each section · the server enforces the lock Section-by-section attestation AI drafts four sections of a clinical note. A provider attests each one. Three are attested and one is not, so signing stays locked at three of four. AI draft History Assessment Plan Orders Sign note locked · 3 / 4 Provider attests each section · the server enforces the lock
Fig. W-02No blanket approval; each section is attested

Shipped by the founding team

AI drafts. A professional attests every section.

AI drafts clinical notes, and signing stays locked until a provider has attested each section. That is human-in-the-loop as a product feature, enforced by the server.

AI agent development

We run agents ourselves, inside hard limits.

Our own engineering system directs AI agents every day, including on this website. Every task runs in an isolated environment, under a spending ceiling the provider enforces rather than one checked afterwards, and a second, independent model reviews the work before it counts as done. An engineer owns every merge. Your agents get the same limits. See how it works.

Timeline and pricing

A priced discovery, then a fixed quote.

New software is quoted, not listed. Discovery is short and has a fixed price agreed before it starts. It ends with a build plan, a schedule and a fixed quote for the build, and you decide whether to go ahead.

What moves the quote: how much is new product versus a feature in existing software, how costly a wrong answer is (and so how much review and evaluation the feature needs), which data sources it must connect to, and where your data is allowed to go. Model usage is billed by the provider; the plan estimates it per request so you know the running cost before launch. Our published prices, for checking, fixing and looking after an app that already exists, are on the pricing page.

Proof

AI the founding team has shipped and runs.

AI-drafted clinical notes with section-by-section attestation and multi-model clinical review are in production in a telehealth platform the founding team built for TrueEval Medical Group, a company the founders are affiliated with. The engineering system behind our own work is built and in daily use. Across the systems the founding team built and runs:

  • 20,900+ automated tests across the systems the founding team built and runs
  • 16,800+ CI runs automated builds and test runs on GitHub Actions
  • 11,000+ commits to production and product codebases

Counted from the repositories and CI history of the systems the founding team built and runs, 22 Sep 2026; each figure rounded down. How we counted

Questions buyers ask

Before you build with AI.

Which AI models do you work with?

We are not tied to one provider. We build with commercial models such as OpenAI's, Anthropic's and Google's, and with open-weight models, and choose per task on measured quality, cost and where your data is allowed to go. We design so a model can be swapped without rewriting the feature.

How do you stop an AI feature from making things up?

Three ways together. We ground answers in your own sources and show where each came from. We measure the feature against a test set before and after every change. And where a wrong answer is costly, a rule, a second model or a person checks the output before anyone relies on it.

What is an AI agent, and do I need one?

An agent is a model that can take steps, such as calling an API, reading a file or sending a message, rather than only answering. Many problems do not need one; a single well-tested model call is cheaper and easier to trust. When an agent is right, we limit its tools and spending and put approvals in front of the actions that matter.

Can you add AI to the app we already have?

Yes. Most of our AI work is inside existing software: a feature added to a product that already has users, data and a release process. We work in your repository, alongside your team if you have one.

Will our data be used to train models?

That depends on the provider and the plan, so we check each provider's terms with you before any of your data reaches it, and we write down which providers see which data.

How long does an AI app take to build?

It depends on the job, so the discovery ends with a schedule rather than a guess. A feature inside an existing app is usually a smaller job than a new AI-native product, and the evaluation work is part of the schedule, not an extra.

What does AI app development cost?

An AI build is quoted, not listed, because its cost is set by what it must do. Discovery has a fixed price agreed before it starts, and the build plan comes with a fixed quote. Model usage is billed by the provider, and we estimate it per request in the plan so you know what running the feature will cost. Our published prices.

What should your product be able to do?

Tell us the feature or product you want AI to power. We will tell you how we would build, measure and ship it.

Discuss your AI project
Next sheet · A‑060 SaaS Development